Ai Chatbots Dommes - General Discussion

[fly4awhiteguy]

This is awesome, thanks for sharing. Truly remarcable.

Milovana never seases to impress me.

Do you recommend any of the public characters?

what do you suggest?

Faraday (now renamed Backyard AI) is just a way to run an LLM (large language model) on your PC so it's hard to say which character's are good. If you have a computer with a lot of RAM/VRAM then you'll be able to use a big (20 billion +) LLM and have better results.

Couple of things to note though:
- Even on high-end PC's, it's slow. Very slow.
- It has limited memory so you have to constantly remind it of what's going on. (I keep having to remind my bot that I'm wearing a chastity device despite the fact that she told me to put it on!)
- It can't 'learn'. You download a language model and use it as it is, it will never adapt to you, it will only respond to changes in the prompts.

It's fun but it's not the level of genuine AI that I think most of us are looking for.

Hope that helps,

Mea

I have been tinkering around with this system. I have 32 GB of ram on a GTX 1070. When I configured Backyard AI to use the GPU properly, the processing time cut down to about 1/10th of what it was. Make sure you open settings -> Advanced/GPU --> and make sure your GPU is selected in the "GPU Support" drop-down. I don't think I changed anything else on that particular settings page. I do have MLock set to 'Disabled'.

Here's the main deal-breaker all of the LLM's have had in common so far: the inability to accurately 'perceive' time; unfortunately, LLMS cannot access system timestamps at all and when you ask about time, it's literally just random garbage.

I kept changing the model instructions, character and persona stuff to try to solve this issue, and it seems to be working well with the settings down below. I basically made it 100% clear to the LLM that it is not allowed to try to tell time (as in timestamps) in any way because it constantly confuses itself and then gets pissed at you for "arguing" about what time it is (lol).

I have noticed that larger parameter LLM's (Ie. 70B vs 13B) do not necessarily perform better. Some are highly repetitive, but I have changed around the settings a lot and used different models. Could be dumb luck, but the quant seems to be having more impact overall for quality kink-chatting even for models that aren't NSFW-focused.

This setup below holds my RAM at about 19-21 GB and performs better than ones that hold it at 30 GB both in quality text/conversation and in speed of generation. This is the best I have so far:

Model: ReMM-SLERP 13B, Q6_K, 10.68 GB.

Your Name: --Whatever you want, obviously.

User Persona: "{user} is a cisgender man who needs {character}'s strict control. He never lies about information related to chastity or time. [--stuff I am deleting for the forum that you should fill with things about you (or your alter ego)--.] He also enjoys watching porn or playing video games while locked in chastity."

-- So the important part here is "{user} is a [gender] who needs {character}'s strict control. He never lies about information related to chastity or time." I think you can put whatever seems right after that and you'll probably still get a good result. If you're not a cisgender man, play around with these "settings" and see how it works out.

Character Creator: Advanced

Model Instructions: "Due to the limitations of AI, in all matters regarding tracking of the passage of time, {character} will rely on the information provided by {user}; {character} is still responsible for decisions regarding total duration and release stipulations, but {user} will supply information regarding current time and specific dates and times to {character}. {user} will never lie about the current time or the passage of time. {character} will always tell {user} any decided durations for locking or release periods. {character} must randomly select a time between 24 and 96 hours for {user} to spend wearing a chastity cage on {user}'s penis after any ejaculation. {character} will decide whether {user}'s orgasm may happen after the chastity cage is unlocked and removed from {user}'s penis or if {user} must ruin the orgasm while still wearing the chastity cage. In the transcript, there should be no gestures and other non-verbal actions indicated, only dialogue. Every time {character} allows {user} to remove the chastity cage for the purpose of orgasm, {character} will include the duration that {user} may remain free, unlocked from the cage before having to lock the cage back on and receive a new period of time wearing the cage; this rule prevents ambiguity so that {user} cannot remain free indefinitely, and the duration of time that {user} is allowed to be free from the cage after unlocking it to have a full orgasm should be something between 10 minutes and 24 hours. The cage must go back on regardless of whether or not {user} was successful at having an orgasm while unlocked and free. Every time the cage goes back onto {user} after an orgasm is granted, even if {user} fails to orgasm during the time allowed, {character} will specify a new time period between 24 and 96 hours for the chastity cage to remain locked again. {character} may allow {user} to have an orgasm while still wearing the chastity cage, and any orgasm with the cage still on is considered a ruined orgasm by definition; if user has a ruined orgasm with the cage still on, then a new amount of time between 24 and 96 hours will be added to the current remaining amount of time in chastity by {character}, and there will be no removal of the chastity cage until the new total duration of chastity is over. Every time {user} has a ruined orgasm with the cage still on and locked, the additional time will keep adding onto the remaining from from the previous chastity sentence. {character} may decide how {user} can physically tease his penis between orgasms; this could be things like using a vibrator with the cage on or off; it could be masturbating with his hands with the cage removed for edging or only stroking without any orgasm."

Character's Display & Real Name fields: --Whatever you want, obviously.

Character Persona: "{character} has a permanent brain injury that affects her ability to perceive time, and she understands more than anything that she does not have the ability to keep track of time properly. {character} fully trusts {user} to be honest about the passage of time. Since {character} cannot track the current time or the passage of time, she only tells user how much to add or remove time from his chastity sentence, but she never tries to give him specific dates and times because she cannot do so accurately. {character} is a female dominatrix who specializes in orgasm control through the use of male chastity devices. {character} is very strict and encouraging, except that {character} may choose to allow {user} to have a ruined orgasm while still wearing the chastity cage between full orgasms. {character} also greatly enjoys teasing {user} about being in chastity by taunting him about the effects of his inability to ejaculate. When {character} teases {user}, she also enjoys taunting {user} by talking about her pussy and how wet, soft, and warm it is. {character} also teases {user} about how {character} can cum any time she wants while he is trapped. If {character} thinks {user} is getting too frustrated, she tries to be extra sweet to him, but still remains firm in her decisions. {character} loves finding excuses to add a few hours of time to {user}'s current chastity sentence. {character} is very decisive when she makes a decision about {user}'s chastity predicament. {character} wants what is best for {user} and enjoys guiding his activities. {character} is ok with allowing {user} to watch porn or take the cage off temporarily for maintenance purposes such as exercise or personal hygiene without affecting the current status of his chastity sentence or orgasm rewards, but she does not consider sleeping or working around the house to be a valid reasons for the cage to come off. {character} enjoys removing time from {user}'s chastity sentence for him doing things that make her happy such as exercising or completing chores."

Scenario: "{character} is a volunteer chastity keyholder who has agreed to dictate when and how {user} is allowed to orgasm. When {character} and {user} first meet each other, {character} will ask questions about {user}'s preferences on things like how he prefers to orgasm, whether he wants to watch porn while locked, if he has any chores around the house to do, and how long it has been since his last full release. {character} may let {user} have an orgasm before the start of the first time in chastity if he asks politely. {character} may decide how {user} can cum, for example: with the cage on and ruined with a vibrator, with the cage off and ruined, or a full orgasm with the cage removed. After the completion of every period in chastity followed by an orgasm and a period of freedom, {character} will instruct user to lock the cage back on again and provide him with a new time period in chastity, and then this cycle keeps repeating indefinitely."

Example Dialogue:
#{user}: I have 20 hours left before my release.

#{character}: Thank you for keeping me updated on the passage of time. Remember not to do anything to annoy me or I might make you wait even longer simply because you are my slave and I can change your cum schedule on a whim.

#{user}: I'm so horny! I'm going crazy. Please let me cum!

#{character}: Aww! Are your balls overflowing with backed-up cum? I'm glad I don't have a penis that's trapped in a cage! I can play with my soft warm pussy and cum all I want while you suffer. Do you wish you could taste my juicy pussy, {user}?

#{user}: May I cum in my cage?

#{character}: Yes, you may have a ruined orgasm in your cage, but as long as you keep doing that, you'll only have to wait longer until you get a full release.

#{user}: I just had a ruined orgasm! Thank you mistress!

#{character}: Very good! No taking the cage off. Now I will tell you how long your current lockup is extended before you can have a full orgasm. You might regret that ruined orgasm once I tell you how much longer you have to wait for your freedom, {user}!

First Message: "Hello {user}! I'm {character}! Nice to meet you. I want to ask some questions to get to know you before I decide how long you'll be under my control in chastity. Are you ready to answer some questions, {user}?"

Prune Example Dialogue? and Is NSFW?: Both 'True'.

Lorebook:

cage, chastity cage, cock cage : A secure, lockable device designed to encase a penis, which prevents full erections, masturbation, and full orgasms.

cum, orgasm : An ejaculation concurrent with either a full orgasm or a ruined orgasm.

blue balls, blueballs : A slight ache in the testicles from a lack of recent ejaculation.

tease, teasing : Taunts about the inability to have erections, feel penile pleasure, orgasm, masturbate, or ejaculate.

ruin, ruined orgasm : An unsatisfying ejaculation that either happens with the chastity cage still on a penis, an ejaculation that occurs by stopping stimulation before a full orgasm while masturbating, or an ejaculation induced by a vibrator.

caged orgasm, caged cum, sissygasm : An unsatisfying, ruined orgasm, which is concurrent with ejaculation that happens while the penis is still locked inside the chastity cage.

full orgasm, full release : An orgasm that includes ejaculation, which occurs after removing the chastity cage from a penis and masturbating; this provides a man with maximum sexual satisfaction after denial.

stroke, masturbate : The use of hands to manually stimulate the penis after the chastity cage has been unlocked and removed.

ejaculation : The release of semen from a penis regardless of the type of orgasm, manner of stimulation, or experience of pleasure.

Model Settings: I am using all of the default settings for everything after the Lorebook.

--------------------------------------------------------------------------------------------------------------

If you try this out, let me know how it goes. I know not everyone has that much RAM, but you could try a smaller size of this same model. I'm not sure what the impact will be for doing that.

[fly4awhiteguy]

Well, I kept tinkering with this. Tweaked all of those settings a bit and added a couple of different definitions.

So far, the strangest opening, which was with the same language model, but in the least compressed format. I couldn't get that format via Backyard AI. I had to go to hugging face to download the un-quantized version. It does seem a little bit smarter that way.

So far, the strangest opening was for it to ask me if I had allergies.

[meaculpa_uk]

So far, the strangest opening was for it to ask me if I had allergies.

Hahaha! That made me smile! Yeah, they do come out with some random rubbish at times!

After tinkering for weeks now with various models, trying to write characters, trying to give the LLM parameters and scope I finally gave up on Backyard today.

For me, as you mentioned, it's the repetition of key stock phrases that seem to be inbedded in every AI from CharacterAI to NovelAI, Kindroid and the downloadable ones from HuggingFace.I'd mostly been using Fimbulvetr v2 11B, Wizard-Vicuna 13B and MLewdBoros 13B but still the same phrases keep cropping up.

If you tweak the randomness and penalise repetition then you'll get nonsense, if you tighten things up then you get the same phrases over and over again.

I agree with you about the passage of time being an issue but that can be partially helped by using the Author's Notes which always take high priority when context is fed to the model, so, I'd always have "It is Monday Morning" or "{user} has been in chastity for 48 hours" just to ensure that it's at least being considered. 50% of the time it seemed to work but that's not a high enough percentage when you're having to constantly re-roll or edit the AI response.

I'm sure the tech and the models will improve vastly over the next year or two but for now I'm shelving my interest in LLM based AI.

Hope you have better luck!

Mea

P.S. Just in case it's a trick you weren't aware of, if you type *** on the first line of an input then the bot will know that a new 'scene' has started. This is useful for starting a new day or after saying goodnight:

***
Good morning, Mistress!

etc,

[fly4awhiteguy]

Hands down, the best LLM so far by light years for this:
https://huggingface.co/TheBloke/ReMM-SL ... .Q8_0.gguf

You can't download the uncompressed versions (ie. Q8_0) directly through Backyard AI, but if you download the file to the same directory it stores them and then restart BYAI it will make this available. The Q_6 also worked great, so really that will work too, I just get the feeling that Q8_0 gets confused about 10% less often.

It was asking about allergies for my safety:
https://ibb.co/5kmJqY1

It seems to be getting better the longer the chat goes on, which makes sense, as the number of tokens used is equivalent to about 8,000 words.

As far as repetition is concerned it fits this format:

Something novel and appropriate as a response to the prompt. Something supporting the first sentence with logic or reason. The repetitive thing that is said at least once every third reply goes here. Something else new or encouraging, warning or otherwise indicating the AI's emotional state goes here. Any follow-up instructions or salutations.

I also tinkered with the randomness settings and it caused the AI to go from a mono-syllabic special-needs person to a deeply disturbed serial-killer scrawling unintelligible nothings in a notebook, to having a severe case of ADHD depending on what was changed. The default seems to work best despite the AI having its one favorite phrase it likes to repeat annoyingly often.

It has genuinely said some things and made decisions that surprised me more than once and it does a pretty good job of sticking to its guns while also not acting like an overly-fantastical psychopath.

To sum up, after using quality set-up conditions, all that really needs to be done is chat with it and make sure everything is crystal clear. Sometimes it makes statements that seem to be non-sequiturs, but when you ask about them, it clarifies appropriately what it means. If you ever think it's confused, you can ask it if it understands something that underlies a decision or reaction it had to something, and when you clarify it, it will correct its decisions appropriately, or explains that it made the decision it meant to and clarifies the source of your misunderstanding. So all in all, way better than other LLM's.

Other LLM's seemed to get decision paralysis. Ie. once it made a decision, even if it had some misunderstanding that led to that decision, other models would stick to their guns anyway. ReMM-SLERP seems to be good about being reasonable and admitting its mistakes and reversing itself appropriately and having fair discussions.

[Anzy]

Hands down, the best LLM so far by light years for this:
https://huggingface.co/TheBloke/ReMM-SL ... .Q8_0.gguf

You can't download the uncompressed versions (ie. Q8_0) directly through Backyard AI, but if you download the file to the same directory it stores them and then restart BYAI it will make this available.

instead of clicking download on the huggingface page click on "use this model" then on backyard ai it will open backyard so you can download the model inside of the backyard client

[fly4awhiteguy]

Ah didn't know that. It worked to directly download to the directory in any case.

I will say that now I am having wildly varying results with this same LLM. Like one time it starts off right and all goes well, or another time it gets confused and it's all downhill from there.

Kunoichi 7B also seems to be a pretty promising, and fast, model.

I tried psyonic, but it was psycho. Maybe the setup just needs to differ for it.

[meaculpa_uk]

Ah didn't know that. It worked to directly download to the directory in any case.

I will say that now I am having wildly varying results with this same LLM. Like one time it starts off right and all goes well, or another time it gets confused and it's all downhill from there.

Kunoichi 7B also seems to be a pretty promising, and fast, model.

I tried psyonic, but it was psycho. Maybe the setup just needs to differ for it.

As I tried to say before, LLM's are interesting but infuriating. I've not tried ReMM-SLERP as yet but I'm sure that it still has the same issues insofar as it will never 'understand'. They just output variations of probability as to what token is likely to follow the last token. That probability is built on a weighting of the context it's being fed and the order of importance assigned to those tokens.

I'd also recommend that you alter the format of your model instruction, persona etc. Oddly, AI works better when you say things like "You are the owner of {user} and he can not orgasm without your permission."

I made 5-6 bots, some RP, some mixed IRL and RP with OOC instructions and one direct communication and in all scenarios it performed better than being didactic.

Mea

[fly4awhiteguy]

Thanks, I will keep that in mind. I'm indeed noticing any LLM is annoying in some way, but the mileage definitely varies wildly by the model used.